Large Language Models hallucinate frequently in critical applications
Researchers studying large language models found they hallucinate—generate false information—far more frequently than commonly understood, with error rates rising significantly as input text length increases. A team at Tsinghua University determined that hallucinations stem from a tiny fraction of neurons formed during initial training, suggesting the problem may be inherent to the technology.
What happened
JV Roig, an expert in agentic AI at Kamiwaza AI, tested different large language models in a controlled environment using input texts of 32,000, 128,000, or 200,000 words and measured how often the models hallucinated when asked questions about the texts.
At 32,000 words, the best-performing model—Zhipu AI's GLM 4.5—hallucinated 1.2% of the time, with an average of 6.8% across models. At 128,000 words, GLM 4.5's error rate rose to 3.2%, while the average jumped to 10%. For models tested with 200,000 words, hallucination rates skyrocketed, with some breaking down entirely.
Reported impact
The research indicates that hallucinations occur far more often than most people realise, and the issue gets worse the more information is provided to the model. In the real world, where models are trained on the entire internet, hallucinations are likely an even bigger problem than shown in the controlled experiment.
Hundreds of billions of dollars are riding on the assumption that artificial intelligence will be reliable enough for high-stakes work, yet the persistent hallucinations suggest this assumption may be unfounded.
Index analysis
In December, a team from Tsinghua University published a paper tracing hallucinations to a tiny fraction—typically less than 0.1%—of an LLM's neurons, the basic decision-making units that process information and pass signals forward.
The problematic neurons are formed during the model's initial training, not at a later stage during fine-tuning, and the study indicates it is impossible to simply identify and remove faulty neurons to fix the problem.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.